This is a register, not an essay. It carries no issue number because it is not fixed at publication — it is revised whenever a figure moves, a definition tightens or a new measurement is made. The version and revision date above are the handle. Essays in Delta Insights are numbered because they are finished; references are versioned because they are not.
We ask this industry to stop grading its own homework. That obligation starts with our own numbers, so this is the piece where we show the working.
Every figure we publish appears below. For each: what it counts, what it does not count, how it is derived, and the honest objection to it. If a number appears on any Empiric Earth material and is not defined here, that is an error.
10B+ real-world miles
What it counts. Cumulative miles of vehicle travel observed by a connected sensor and processed into the record, across the combined company, since 2015.
What it does not count. Estimated miles, simulated miles, or miles extrapolated from a sample. Every mile in the figure was observed.
The honest objection. Volume is the least interesting thing about a record. Ten billion miles of uneventful motorway is a large number and a small amount of information. We publish it because it is the fastest way to convey the scale of elapsed exposure, not because it is the most meaningful figure here.
60M+ edge cases
What it counts. Events in the record that satisfy all three of the conditions we use to define an edge case: rare in frequency, high in potential consequence, and poorly represented in average-distribution datasets. Compound events, where two or more ordinary factors occur together, are counted once.
What it does not count. Every anomaly, every hard braking event, or every alert. The threshold is consequence, not deviation from the mean.
The honest objection. The boundary is a judgement. Draw it generously and the number grows without anything changing in the world. That is exactly why the definition is published here and why we do not lead with this figure. The defence of the number is the definition, not the magnitude.
300M+ new miles per month
What it counts. New observed miles per month across the combined company, all networks and products.
The honest objection. It is frequently confused with the Atlas figure below. They measure different populations, and quoting either as the other inflates both.
100M+ new Atlas miles per month
What it counts. New observed miles per month flowing through the Atlas network specifically, which is a subset of the combined company total.
Why both exist. Atlas is the retrieval product. A researcher asking what is queryable needs the Atlas figure. An investor asking about the company’s observational footprint needs the combined figure. Publishing only one would mislead one of them.
50+ countries
What it counts. Countries in which the combined record contains observed miles.
The honest objection. Presence is not depth. Coverage in a country ranges from dense metropolitan observation to a thin corridor. We do not publish a per-country depth figure yet, and until we do, this number should be read as reach rather than as capability everywhere.
350K+ connected sensors · 1.2B+ miles indexed annually
What they count. Devices actively contributing to the Atlas network, and miles indexed into it over a rolling year. Both are Atlas-scoped.
2× reaction time · 3–5 seconds earlier · 99% alert accuracy
What they measure. Collision-alert performance under controlled testing by the Virginia Tech Transportation Institute. Reaction time is measured against a driver receiving a conventional warning. Lead time is measured against conventional forward and pedestrian collision warning. Accuracy is the proportion of alerts corresponding to a genuine developing conflict.
What they are not. A benchmark of any perception model, including ours. VTTI tested alerting. Merging these figures with a model benchmark produces a claim neither supports, and it is the single most common error made with our numbers — including, occasionally, by us.
67% national logistics fleet · 62% bpx energy · 68% KeHE Distributors
What they count. 67% is the reduction in most-serious collisions — tow-aways, hospitalisations and fatalities — at a national logistics fleet of 34,000 vehicles across 600 locations and 70,000 drivers, measured from pilot to fleet-wide over 24 months, with comparable fleets on the same FMCSA data flat over the period. 62% is at-fault collisions at bpx energy. 68% is total collisions at KeHE Distributors. Each is against that customer’s own baseline.
Why they differ. One is most-serious only, one is at-fault, one is total. Three different measures. The spread between 62 and 68 says nothing about which deployment worked better, and reading it that way is the most common mistake made with these figures.
The honest objection. These are customer-scoped and are not evidence of a universal effect. Fleet profile, route mix, baseline period and reporting regime all move the result. A single blended percentage across all deployments would need a method we have not published, so we do not publish one.
What we deliberately do not publish
- A universal collision-reduction figure. It would require a comparator and method across unlike deployments. Until that exists, named customers only.
- Benchmark leadership stated broadly. Our model leads its own published long-tail categories. That is a real and checkable claim, and it is smaller than the one people want us to make.
- Anything described as anonymised. We describe what we actually do to the material instead, because the stronger word implies a guarantee no camera-based system can honestly give.
The rule behind all of it
Δ goes in front of a figure only where that figure has approved scope and method. Never in front of an aspiration, an illustration, or a number we have not defined.
If a number on any Empiric Earth material does not appear in this piece with a definition, treat that as an error and tell us. We would rather correct it here than defend it later.